Method for generating plane creative pictures in batches
Through a generative large model, the text or pictures entered by users are processed and creative graphic images are generated in batches, solving the problem of low efficiency during manual design and achieving a fast and efficient design process.
Patent Information
- Application Number
- CN202411418484.1
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-10-11
- Publication Date
- 2025-02-28
AI Technical Summary
When designing creative graphic pictures manually, it takes a lot of time and effort to conceive, draw and adjust, resulting in slow design processes and low efficiency.
Through user text or picture input, use a generative large-scale model to generate graphic creative pictures that meet the target requirements in batches, including text encoder converting text requirements into vector representation, scene division and generated models to generate pictures.
The design process of creative graphic images has been accelerated, the generation efficiency has been improved, and the time and energy investment of users in conception, drawing and adjustment has been reduced.
Smart Images
Figure CN119540378A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of picture processing, and in particular to a method for batch generating planar creative pictures based on an AIGC generative large model. Background Art
[0002] When completing graphic creative images, designers use colors, graphics, typography and artistic techniques to create attractive and expressive graphic ideas. However, in the manual design process, it takes a lot of time and energy to conceive, draw and adjust, which leads to a slow design process and low efficiency for the entire graphic creative image. In order to solve the problem that in the manual design process of graphic creative images, it takes a lot of time and energy to conceive, draw and adjust, which leads to a slow design process and low efficiency for the entire graphic creative image. Summary of the invention
[0003] By inputting text or pictures from the user, the user's target needs can be determined, and target graphic creative images that meet the target requirements can be generated in batches. Therefore, in the present invention, when generating graphic creative images, the user does not need to spend a lot of time and energy to conceive, draw and adjust, which speeds up the design process of the entire graphic creative image and improves the efficiency of generating graphic creative images.
[0004] A method for batch generating flat creative pictures comprises the following steps: (1) Obtaining user input requirements, where the user inputs text requirements for the subject of an image or an image containing a subject image; (II) When the user inputs text, it includes: 1. Use the text encoder to convert the text requirements into vector representation, obtain the numerical expression of the text requirements, and extract the style feature information and emotional color features implicit in the Chinese characters; 2. Divide the text requirements into scenarios according to the style feature information and emotional color features, and obtain multiple scenario definitions, which are different from each other; 3. Generate multiple target plane creative pictures that meet user needs based on multiple scene definitions through a generative model; (iii) when a user inputs a picture requirement, extract the subject features including the subject in the input picture, and generate multiple flat creative pictures including the subject through the generative model; (iv) Before steps (ii) and (iii), perform model training to obtain a generative model: 1. Obtain the data set for model training and search the database for database images corresponding to multiple scene definitions; 2. Preprocessing the multiple database images by cleaning, cropping, size standardization, and format conversion; 3. Use adversarial diffusion distillation technology to train the model, that is, introduce noise sequences and progressive diffusion processes in the training process to extract high-quality sample data from the training data set to improve the training effect of the generative model, or use three-order cascade diffusion technology to train the model, that is, design the generative model as a model structure containing three cascade diffusion networks. The three diffusion networks gradually generate higher quality images, and use three-order cascade diffusion technology to train the model to achieve the gradual generation of high-quality images.
[0005] Also includes: (5) after generating the plurality of target plane creative images, performing quality assessment on the plurality of generated target plane creative images to assess the clarity and authenticity of the generated target plane creative images; (vi) performing post-processing operations such as denoising, enhancing, and cropping on multiple target graphic creative images to ensure that the generated target graphic creative images meet expectations and meet the required quality standards; (VII) Displaying multiple graphic creative images on a user interface, allowing users to manually evaluate and select the generated graphic creative images, and then adjusting the parameters of the target generative model based on user feedback.
[0006] Beneficial Effects The target text is parsed, and the scene of the theme is split according to the parsing results to obtain multiple scene definitions. Different scene definitions can be used to convey the same theme, that is, there can be multiple different scene definitions under the same target theme to interpret and express the style and emotion corresponding to the target theme. The user enters a theme name and can obtain multiple graphic creative images in completely different scenes that interpret the theme. When generating graphic creative images, users do not need to spend a lot of time and energy to conceive, draw and adjust, which speeds up the design process of the entire graphic creative image and improves the efficiency of generating graphic creative images.
[0007] By generating creative graphic images through generative models, the characteristics and distribution of the data can be fully learned, and images that are significantly different from the data set can be generated, effectively managing copyright. BRIEF DESCRIPTION OF THE DRAWINGS
[0008] Figure 1 The technical architecture of the image generation method; Figure 2 The overall generation process of graphic creative images; Figure 3 Examples of training and subject training. DETAILED DESCRIPTION
[0009] The present invention provides a method for batch generating graphic creative images to solve the problem of low efficiency in the process of manually designing graphic creative images. The specific scheme is as follows: (1) obtaining a user input requirement, where the requirement may include at least one of the following: a text requirement of a desired picture theme or a picture containing a subject image input by the user; (II) When the user inputs a text requirement, it includes: 1. Use the text encoder to convert the text requirements into vector representation, obtain the numerical expression of the text requirements, further analyze and interpret the vector representation, and extract the style feature information and emotional color features implicit in the Chinese characters; 2. Divide the text requirements into scenarios based on style feature information and emotional color features to obtain multiple scenario definitions. These multiple scenario definitions are different from each other, but all convey the theme input by the user to interpret and express the style and emotion corresponding to the theme; 3. Generate multiple target plane creative pictures that meet user needs based on multiple scene definitions through a generative model; (III) When the user inputs a picture requirement, the main features including the subject in the input picture are extracted, and multiple flat creative pictures including the subject are generated through the generative model.
[0010] (iv) Before steps (ii) and (iii), model training is required to obtain a generative model: 1. Obtain the data set for model training and search the database for database images corresponding to multiple scene definitions; 2. Preprocessing the multiple database images by cleaning, cropping, size standardization, format conversion, etc.; 3. Use adversarial diffusion distillation technology to train the model, that is, introduce noise sequences and progressive diffusion processes during the training process to extract high-quality sample data from the training data set to improve the training effect of the generative model. Alternatively, use three-order cascade diffusion technology to train the model, that is, design the generative model as a model structure containing three cascade diffusion networks. The three diffusion networks gradually generate higher quality images. Use three-order cascade diffusion technology to train the model to achieve the gradual generation of high-quality images, which helps to improve the performance and generation effect of the generative model.
[0011] (5) after generating the plurality of target plane creative images, performing quality assessment on the plurality of generated target plane creative images to assess the clarity and authenticity of the generated target plane creative images; (6) Performing post-processing operations such as denoising, enhancement, and cropping on multiple target graphic creative images to ensure that the generated target graphic creative images meet expectations and meet the required quality standards.
[0012] (VII) Displaying multiple graphic creative images on a user interface, allowing users to manually evaluate and select the generated graphic creative images, and then adjusting the parameters of the target generative model based on user feedback.
[0013] Example: (1) Obtaining the text theme of "Working People" input by the user; (ii) Convert the text of “dagongren” into numerical representation, conduct further analysis and interpretation, and extract the style feature information and emotional color features implicit in the Chinese characters; (3) Based on the style feature information and emotional color characteristics implied in "working people", the theme "working people" is divided into multiple scene definitions such as "typing on the keyboard", "moving bricks", "delivering food", and "presenting PPT"; (iv) Generate multiple target plane creative images that meet user needs based on multiple scene definitions through a generative model; (VI) Obtaining the image requirements input by the user, it is necessary to include the company's IP image in the generated image, identify the facial features and morphological features of the IP image, apply the above facial features and morphological features to different scenarios, and generate multiple flat creative images through a generative model.
[0014] (VII) Text requirements and image requirements are input simultaneously, and the theme "Working People" is split into multiple scene definitions such as "typing on the keyboard", "moving bricks", "delivering food", and "presenting PPT", etc. The facial features and morphological features of the IP image are identified, and the above facial features and morphological features are applied to the above multiple scene definitions. Multiple graphic creative images are generated through the generative model, and each generated graphic creative image contains one or more of the above N scenes and includes the company's IP image.
[0015] (VIII) According to a plurality of scene definitions, a plurality of target plane creative images are grouped into a plurality of groups of target plane creative images, and each group of target plane creative images is grouped and coded, and preset entries corresponding to the plurality of scene definitions and the group codes corresponding to the plurality of groups of target plane creative images are displayed on a user interface, and a user can select a preset entry on the user interface and call the corresponding group code to retrieve the target plane creative image in the group of target plane creative images.
[0016] (IX) Performing quality assessment on the multiple generated graphic creative images to evaluate the clarity and authenticity of the generated target graphic creative images; performing post-processing operations such as denoising, enhancement, and cropping on the multiple target graphic creative images to ensure that the generated target graphic creative images meet expectations and meet the required quality standards.
[0017] (X) The data collection phase of the model training process involves obtaining graphic design works of different styles, such as posters, business cards, brand logos, etc. from sources such as design websites, design communities, and public datasets. The collected design works are cleaned, cropped, size-standardized, and format-converted to ensure that they are suitable for model training. Select appropriate generative image models, such as DCGAN, StyleGAN, etc., and implement them using deep learning frameworks (such as TensorFlow, PyTorch). Use the preprocessed design data to train the selected model so that the model can learn different types of graphic design styles and features. During the training process, monitor the quality and diversity of the generated results, adjust the model's hyperparameters (such as learning rate, batch size, etc.) and the weights of the loss function to optimize the generated designs. Perform preliminary automated evaluation of the generated designs, such as by calculating image quality indicators (such as SSIM, PSNR) to evaluate the clarity and authenticity of the designs. Provide a user interface for users to manually evaluate and select the generated designs, and adjust the parameters of the design generation based on the feedback.
[0018] Glossary Graphic creative images usually refer to creative and artistic images designed and produced on a two-dimensional plane. Such images are usually used in design fields such as advertising, promotion, illustrations, posters, and business cards.
[0019] Dataset: A dataset is a collection of organized data, which is usually used in machine learning, deep learning, data analysis and other fields. A dataset can contain various types of data, such as text, images, videos, audio, etc., as well as the labels or annotation information corresponding to these data.
[0020] Generative Model: A generative model is a type of machine learning model that aims to simulate or generate new data that is similar to training data. In contrast to the discriminative model, the generative model can not only classify or predict data, but also learn the distribution characteristics of the data from the training data, and then use these characteristics to generate new data samples. Generative models are widely used in many fields, including image generation, natural language processing, audio synthesis, etc.
[0021] Image noise: refers to random or irregular pixels or pixel values that are not expected to appear in an image, usually caused by interference or distortion introduced during image acquisition, transmission, processing, etc. These noise points may appear in the form of brightness changes, color deviation, block distortion, spots, etc., which will reduce the quality and clarity of the image, affecting the viewing experience and the effect of subsequent image processing.
[0022] Three-stage cascade diffusion: is a deep generative model technique for image generation, often used to generate high-quality, high-resolution images. This technique is based on a diffusion process that gradually generates images through multiple cascaded steps, each with its own generative network.
[0023] Text Encoder: refers to a model or algorithm used to convert text data into numerical representation. In the field of natural language processing (NLP), text usually exists in the form of words or characters, and computers cannot process text data directly, so text needs to be converted into numerical representation for subsequent processing and analysis.
[0024] Adversarial Diffusion Distillation: is a deep learning technique used to train generative models to generate high-quality, high-resolution images. This technique combines the ideas of adversarial training and diffusion distillation to improve the stability and generative ability of generative models.
[0025] Multimodal model control: refers to a model that can process multiple different types of input data (such as text, images, speech, etc.), and can make corresponding decisions or outputs based on different types of input data. This type of model control has a wide range of applications in the field of artificial intelligence, especially in tasks such as cross-modal learning and multimodal fusion.
Claims
1. A method for batch generating creative graphic images, characterized in that The steps include: (1) Obtaining user input requirements, where the user inputs text requirements for the subject of an image or an image containing a subject image; (II) When the user inputs text, it includes:
1. Use the text encoder to convert the text requirements into vector representation, obtain the numerical expression of the text requirements, and extract the style feature information and emotional color features implicit in the Chinese characters; 2. Divide the text requirements into scenarios according to the style feature information and emotional color features, and obtain multiple scenario definitions, which are different from each other; 3. Generate multiple target plane creative pictures that meet user needs based on multiple scene definitions through a generative model; (iii) when a user inputs a picture requirement, extract the subject features including the subject in the input picture, and generate multiple flat creative pictures including the subject through the generative model; (iv) Before steps (ii) and (iii), perform model training to obtain a generative model:
1. Obtain the data set for model training and search the database for database images corresponding to multiple scene definitions; 2. Preprocessing the multiple database images by cleaning, cropping, size standardization, and format conversion; 3. Use adversarial diffusion distillation technology to train the model, that is, introduce noise sequences and progressive diffusion processes in the training process to extract high-quality sample data from the training data set to improve the training effect of the generative model, or use three-order cascade diffusion technology to train the model, that is, design the generative model as a model structure containing three cascade diffusion networks. The three diffusion networks gradually generate higher quality images, and use three-order cascade diffusion technology to train the model to achieve the gradual generation of high-quality images.
2. The method for batch generating creative graphic images according to claim 1, characterized in that Also includes: (5) after generating the plurality of target plane creative images, performing quality assessment on the plurality of generated target plane creative images to assess the clarity and authenticity of the generated target plane creative images; (vi) performing post-processing operations such as denoising, enhancing, and cropping on multiple target graphic creative images to ensure that the generated target graphic creative images meet expectations and meet the required quality standards; (VII) Displaying multiple graphic creative images on a user interface, allowing users to manually evaluate and select the generated graphic creative images, and then adjusting the parameters of the target generative model based on user feedback.